Improved Remote Sensing Image Rotating Target Detection Algorithm Based on Transformer

Abstract

As satellite remote sensing and aerial photography technologies continue to advance in recent years, there has been a noticeable increase in both the resolution and image quality of remote sensing images. Furthermore, an abundance of data sources has emerged, intensifying the challenges associated with detection. To address the challenges posed by small object size and dense distribution in remote sensing images, an innovative solution has been introduced. This solution entails an enhanced rotating object detection algorithm which leverages the power of vision Transformer technology. By utilizing this approach, the aim is to overcome the limitations of poor robustness and low detection accuracy commonly encountered in such scenarios.The enhancement of the feature extraction capability of the detection algorithm in YOLOv4’s feature fusion part is achieved through the introduction of the MS-Transformer module. This module, known for its self-attention mechanism, facilitates the acquisition of pertinent information among targets, thereby bolstering the algorithm’s ability to detect densely distributed targets. Moreover, the advancement of the five-coordinate YOLOv4 object detection framework enables the realization of multi-angle remote sensing object detection. To mitigate the issue of overlapping prediction frames on dense targets, the model incorporates the soft-NMS suppression method, ultimately refining the detection performance. The efficacy of the proposed algorithm in improving the model’s detection capability is substantiated through experimentation using the DOTA dataset.